1.3 KiB
name, description, disable-model-invocation, metadata
| name | description | disable-model-invocation | metadata | ||||
|---|---|---|---|---|---|---|---|
| writing-for-agents | Write assigned agent-facing repository guidance with precise triggers, scope, and verification. | true |
|
Writing for agents
Use when the Plan assigns an edit to a repository skill, AGENTS.md, or another agent-facing document. Inspect the existing instruction hierarchy and the target document's readers before editing. Keep each pointer short: state what document it leads to and the distinct conditions that require reading it. Put detailed instructions in the target document rather than repeating them in every pointer.
State the agent's allowed decision, required inputs, observable output, and escalation condition. Use concrete paths and examples only when they are stable in the assigned repository. Keep rules at the narrowest scope that reaches every intended reader; avoid contradictory duplicates. Check that each imperative can be followed with the role's actual tools and permissions. For RelayView skill packages, keep the package instruction-only and single-role; never use text to grant tools, credentials, workflow transitions, or another role's authority. Run the relevant validation and report any instruction conflict to the owner.